Gemini’s API can now hold a tool’s parameter schema back until the model asks for it
Lucia Loher and Nora Kassner’s post is dated 19 November. Tool Retrieval is a preview in the Interactions API: deferred tools keep a name and a description in the prompt, and the parameter schema loads only when the model requests it, inside the same call. The blog post says there is no separate charge. The pricing page says each server-side search is an extra model call, billed as input tokens. The fixed block of tokens is not given a size.
Feature2 minIn-depth
James Okonkwo2 min read
- Gemini API
- Agents
- Tool use

Photo: GoogleSource page
Google is letting a Gemini agent keep a large tool library on standby and load a parameter schema only when a call needs it. Lucia Loher, a product manager for the Gemini API, and Nora Kassner, a research scientist at Google DeepMind, posted the feature on 19 November. Tool Retrieval is in preview in the Interactions API. The problem they describe is mechanical. An agent wired to dozens or hundreds of tools sends every full parameter schema on every request. That, they write, raises the input-token count, adds latency, and makes tool selection less accurate as the list grows.
The switch is a flag and an extra tool. A developer marks a function with defer_loading set to true and adds a tool whose type is tool_search. The model still sees the name and the description of each deferred function. The parameter schema stays out of the prompt until the model asks for that function by name. The API then loads the schema, and the model emits the function call, inside the same request. The post’s example uses gemini-4-hydrogen and a function called issue_refund. The steps it prints are a tool_search_call, a tool_search_result, and then the function_call. There is no index for the developer to host. Gemini runs the search on the server. Tools that are used on most turns can stay loaded. The rest can wait. A remote MCP server can take the same flag, so every tool that server exposes loads on demand. A developer who wants to run the lookup sets the tool_search execution to client. That path adds a round trip, because the application does the search, and it can return a tool that was not declared when the conversation started.
What the two pricing lines actually say
The blog post says Tool Retrieval has no separate charge. Input and output tokens are billed at the ordinary rates for the model on the request. Turning tool_search on adds what the authors call a small, fixed block of input tokens. They do not say how many. Each schema the model loads is then billed as input for the rest of the conversation, on the same terms as a tool that was declared at the start. The pricing page, in the tools table, adds a step. Each server-side tool search runs an extra model call. The prompt for that call is reported as tool-use tokens and charged at the input rate. The savings, the blog post says, show up when the schemas that never load outweigh the ones that do, and it tells developers to measure that on their own traffic. The models named in the post, and again in the 19 November release notes, are Gemini 3.6 and later, including every Gemini 4 model, and only through the Interactions API. The release notes mark the feature as a preview.


